Learning Optimal Lattice Vector Quantizers for End-to-end Neural Image Compression

Neural Information Processing Systems 

This design decision is apparently made in favor of the simplicity and computational efficiency.